Simon A Hewat

anti skeumorph, chief bit shuffler @ stealth, ex ML @scale_AI, @tesla

[Simon A Hewat @sahewat]

@tautologer 1. slow (er) takeoff: im more worried about abuse from institutions than the public. equal access helps here
2. fast takeoff: RSI is going to need a lot of compute and energy. why fight over a tiny planet when the universe is large and silicon can go fast without squishing.

[Simon A Hewat @sahewat]

graduate descent remains undefeated

[Härkönen @yokaihainen]

Researchers at the University of Helsinki are building an immense manually operated data center. Professor Jäätikäinen explains: "Every bit of the 2b parameter LLM is represented by a physical switch. We'll have a man inside, so to speak, and actually literally, namely Tarmo, my PhD student, who will be flipping them according to instructions. He'll finally be able to tell us what it feels like to be AI and whether they're conscious." Critics accuse the project of a horrible misreading of the Chinese Room thought experiment

[Simon A Hewat @sahewat]

@ludwigABAP kalman filter kalman filter kalman filter

[Simon A Hewat @sahewat]

@ludwigABAP semseg is still not solved. SAM is def impressive but still needs much of the traditional vision stack.

you’re also relying on the judgement of the model on what you detect in the first place which… is highly questionable outside the most basic cases (personal, car, chair)

[Simon A Hewat @sahewat]

@ludwigABAP it’s got really cool demo videos!

[Simon A Hewat @sahewat]

@bigsofti @wordgrammer @snwy_me

[Simon A Hewat @sahewat]

@stevenkplus1 @DevinAI @cognition @ScottWu46 incredible

[Simon A Hewat @sahewat]

@LayrKits @SaidAitmbarek dm me

[Simon A Hewat @sahewat]

@tszzl @repligate @genalewislaw pls provide a goblin-mode option in codex

[Jon Barron @jon_barron]

"World Models" discourse will now be paused, pending the invention of terraforming. If we had named LLMs "Thought Models" we'd never get past the philosophical debates around what is *actually* happening under the hood. Just name your model according to its inputs or outputs.

[Simon A Hewat @sahewat]

@keenanisalive @jon_barron as an aside, i think we’re in an unfortunate regime where marketing of techniques has to become progressively more bombastic to justify investment/scale. NVL72s are expensive.

[Simon A Hewat @sahewat]

@keenanisalive @jon_barron i find the process of rastering out view trajectories from these things hilarious but there has to be a better way

[Simon A Hewat @sahewat]

@jon_barron @keenanisalive most of these 3d scene generators use video models anyway, no?

marble + the hosted service today have occlusion/depth artifacts that make me think they do this too.

[Simon A Hewat @sahewat]

@keenanisalive @iquilezles the point cloud isn’t the interesting part to me. it’s the fact that we’re effectively running SLAM on a hallucinated video

[Simon A Hewat @sahewat]

@keenanisalive The model du jour is a video DiT with an additional geometric prior + a RGB-D SFM pipeline tuned for it.

I’m enjoying seeing the bitter lesson briefly unwound as people try to find an end to end representation that plays nice. It’s soothing and nostalgic.

[Simon A Hewat @sahewat]

@RishiRajas28936 @snwy_me that’s actually a pretty good explanation of the chinese room

[Simon A Hewat @sahewat]

@blackbirdsage @snwy_me @joaomendoncaaaa i’d also recommend learning a little math. it’s going to make the new world we’re entering a little less scary

[Simon A Hewat @sahewat]

@blackbirdsage @snwy_me @joaomendoncaaaa @blackbirdsage it’s ok not to have opinions on something you don’t understand.

it’s not ok to justify your ignorance by attacking someone else making a good faith attempt
to help you

[Simon A Hewat @sahewat]

@ZGojcic Is the plan for the code to be released on github today?